A deep learning framework for neuroscience

A deep learning framework for neuroscience
B. Richards et al. 2019

Presented by
Minseok Kang

Contents

1. Systems neuroscience
2. Deep learning
3. DNN as a model of brain
4. Three components of DNN
5. Advantage of the proposed framework
6. Caveats & Discussions

Systems neuroscience

- Traditionally uses local recordings to figure out function of individual neurons which are then combined to form a small circuit

e.g. how retina computes motion [2]
two different bipolar cell types connect with time lag to starburst amacrine cells

- drawback: doesn't scale up to a larger network which possibly has multiple functions
the function of each neuron can't be described in a concise manner
-> using artificial neural network (ANN) as a model

Deep neural network (DNN)

- DNN: hierarchical network with multiple layers of units which are abstract neurons

- Instead of specifying what the function computes, specify three components so that the function can be learned: learning goals, learning rules, architectures

- Currently very successful in varieties of tasks that has to be solved by animals as well
e.g. image/speech classfication/generation, game playing

- Credit assignment problem: how each parameter (weight) should change to assure increase in objective function
-> Often solved using backpropagation, which is considered biologically implausible due to symmetric feedback weights and distinct feedfoward/feedback information passes

DNN as a model of brain

- Similar representational transformation in early perceptual systems [17,28]

- many known behavioral/neurophysiological phenomena e.g. grid cell tuning and visual illusions can be found in DNN as well if trained with similar inputs [24,30,31,32,33]

- Biological implausiblility of backprop is often overstated, as there are methods to estimate gradients without using it explicitly [12,14,34-39]


Richards et al., 2019

Three core components of DNN - Architectures

- Arrangements of units
- Decides inductive biases of the network specific to the problem e.g. CNN

- Empirical studies
    1. Anatomy, i.e. cell types/connectivities and developmental processes
    2. Checking representational similarities between DNN layers/brain regions [8,71]




Three core components of DNN - Learning rules

- How to update parameters

- Importantly, need to address the credit assignment problem
    1. Apical dendritic model [12,14]
    2. Attention-based model [37,38]
- Infer plasticity rules from how representations change during learning [77]



Three core components of DNN - Objective function

- May not be directly observable and in reality, brain probably optimizes multiple fuctions
-> Without the objective function learning won't be learning but just a random change
-> Can be defined without reference to the exact task e.g. predictive coding

- Relating to empirical studies
    1. ethology provides insight to which functions animals should optimize
    2. relating known plasticity rules to potential objective functions [59,82]
    3.  inverse reinforcement learning [83]
    4. correlations between representational geometries in model and in reality [28,84]

Advantage of optimization based models

- Better suited to building a complex model solving real-world problems than building models from individual neural responses
-> Environment is diverse and dynamic which results in responses of individual neurons which are hard to interpret

- Specific computations don't have to be stated
-> More compressed information which are passed through genomes [48]


Caveats

- We don't have to get rid of explanations of response of individual neurons e.g. orientation selectivity
-> rather, optimization-based models can act as a high-level model that provides a simple explanation which are generally applicable like evolution theory in biology
-> this is even used in designs e.g. CNN

- Many behaviors of animals don't require learning
-> brain have strong inductive biases acquired on evolutionary time scale

- ANNs have too many parameters
-> Counterintuitively, over-parametrized learning systems might be good in generalization [42,89]

Discussions

- Do you think the framework introduced here is well justified?

- Are there any alternative general top-down theory?

- If brain optimizes multiple functions, how does it optimize importance of those functions?

Resources

[1] The paper: Richards, B. A. et al. (2019). A deep learning framework for neuroscience. Nature Neuroscience. https://doi.org/10.1038/s41593-019-0520-2
[2] podcast about apical dendrite model: https://braininspired.co/podcast/9/
[3] cognitive computational neuroscience: https://ccneuro.org/2020/

* All references made in the slide are reference numbers in the original article